Vous êtes sur la page 1sur 4

S C I E N C E S C O M PA S S

65 64 63 62 61 60 59 58 57 56 55 54 53
BOOKS: DOING SCIENCE

BOOKS ET AL.
nomena as well as fundamental philosophical questions. Each of these ideas warrants careful consideration.

Is the Universe a Universal Computer?


Melanie Mitchell

52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10

CREDIT: STEPHEN WOLFRAM

istory has seen the development of many new sciences but very few new kinds of science. New kinds of science involve radical changes in thinking, such as the shift from Aristotelian traditions to experimental methods A New Kind and the description of Science of natural phenomby Stephen Wolfram ena in mathematiWolfram Media, Chamcal termsrevolupaign, IL, 2002. 1280 pp. tions associated $44.95, 40, C$69.95. with names like ISBN 1-57955-008-8. Galileo and Newton. Thus it is with no small risk of hubris that Stephen Wolfram titles his account of his approach to explaining the natural world A New Kind of Science. At over 1200 pages, his book rivals the combined lengths of Galileos Dialogues and Newtons Principia . But does it fulfill the promise of its title and heft? Not very well. The books central idea is that the simple computer programs, namely cellular automata, can explain natural phenomena that have so far eluded traditional mathematical approaches such as differential equations. A cellular automaton, in its simplest incarnation, is a one-dimensional line of sites or cells, each of which is either black or white. The color (state) of the cell can change over time. At each discrete time step, every cell updates its stateeither retaining or flipping its color from the previous stepas a function of its previous state and those of its two nearest neighbors. The cellular automaton rule comprises the list of functions that map each three-cell neighborhood to the update state for the center cell. In his theoretical work, Wolfram typically considers the lines of cells limitless to ensure there is no ambiguity at the boundaries. Such one-dimensional, two-state, two-neighbor cellular automata are called elementary; more complicated versions can have additional states per cell, larger neighborhoods to determine update states, and additional dimensions. Cellular automata are perhaps the most idealized models of complex systems: they consist of large numbers of simple components (here, cells) with no central con-

The author is at the Department of Computer Science and Engineering, OGI School of Science and Engineering, Oregon Health and Sciences University, Beaverton, OR 97006, USA. E-mail: mm@cse.ogi.edu

Complexity from Simple Programs Throughout the book, Wolfram presents many beautiful, striking pictures of the betroller and limited communication among havior of various cellular automata. These components. Originally defined and stud- demonstrate that even elementary cellular ied by the mathematicians Stanislaw Ulam automata can produce patterns ranging and John von Neumann, cellular automata from simple to quite complicated, highly have been used as models for such natural ordered to seemingly random. The great phenomena as earthquakes, turbulent flow, diversity of these patterns and the fact that biological pigmentation, and tumor growth. such simple rules can produce such apparThey have also been applied in computer ently complex behavior is viewed by Wolscience as idealizations of massively paral- fram as deeply significant. lel, non-centralized computation. That simple rules can produce complex Wolfram grounds his approach on six behavior is a very important idea, and it unprincipal claims: Simple programs (i.e., ele- derlies the science of complex systems. mentary cellular automata) can produce However, Wolfram implies that the notion highly complex and random-looking behav- was his discovery and the field of complex systems his inventionan absurd claim. The idea of simple rules leading to complex behavior underlies much of dynamical systems theory and particularly the subset often known as chaos theory. I dont know when the idea was first articulated, Image not but by the early 1970s available for Nicholas Metropolis, Paul Stein, and Myron Stein online use. ( 1 ) had provided a detailed explanation of the complex behavior of simple iterated maps (such as the famous logistic map). In the late 1960s, John Conway developed his Game of Life, a simple two-dimensional cellular automaton capable of highly complex beWidth doublers. These examples of three-color cellular automata havior ( 2 ). Around the that achieve the purpose of doubling the width of their initial con- same time, Aristid Lindition were taken from the 4277 cases found in an exhaustive denmayer invented what search of all of the more than 7.625 x 1012 possible rules. are now called L-systems, simple rules that give rise ior. Such programs, implemented in natural to extremely lifelike pictures of plants and systems, give rise to most of the complexity other natural forms (3). One of Wolframs and randomness that we observe in nature. own early contributions was observing and These programs lead to better models of classifying such complex behavior in elecomplex systems than do traditional mathe- mentary cellular automata. matical approaches. Computation in cellular automata and similar simple programs Origins of Complexity in Nature provide a new framework for understanding In Wolframs view, cellular automata or simicomplex systems. Elementary cellular au- lar simple rules are responsible for most of tomata can exhibit the ability to perform the randomness and complexity seen in naany computable procedure. This computa- ture. To support this claim, he notes that rantional universality gives rise to the principle domness and complexity are very common of computational equivalence, which in the behavior of simple rules and that some Wolfram claims is a new law of nature that complex and random-looking phenomena in illuminates many aspects of natural phe- nature have visual features similar to those
SCIENCE VOL 298 4 OCTOBER 2002

www.sciencemag.org

65

S C I E N C E S C O M PA S S
65 64 63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 produced by simple cellular automata. He offers an important and useful discussion of the terms random and complex. The meaning of both terms depends on the features of behavior that are of interest and the available perceptual and analytical tools. Wolfram examines these connections in considerable detail, and he devotes an entire chapter to a discussion of perception and analysis. Nonetheless, there is little support for his claim that because simple rules commonly produce random and complexlooking behavior, simple rules must be the source of most such behavior in nature. In addition, Wolfram distances his results on cellular automata from earlier results on chaotic systems. He contends that whereas some cellular automata can generate randomness intrinsically (starting from a very simple initial condition), a chaotic system can do so only when its initial condition is random. This is incorrect; there are many chaotic systems in which even a very simple initial condition produces a random-looking outputthe Lorenz system of equations (4) is one well-known example. In general, cellular automata are a type of discrete dynamical system and can exhibit behavior analogous to chaos. Thus, many of the results from the theory of dynamical systems apply to cellular automata. Cellular Automata as Models The author traces the origins of his new kind of science to his frustration with analytical approaches. He claims that, in contrast to traditional mathematics, the research program he develops in the book is for the first time able to make meaningful statements about even immensely complex behavior. This promise is not borne out. One chapter of the book presents patterns formed by cellular automata and similar systems that model crystal growth (particularly snowflakes), material fracture, fluid turbulence, plant morphologies, and pigmentation patterns of shells and animals. In each case, the behavior of the model visually resembles in some way the behavior of the actual system; this is particularly striking for snowflakes and sea shells. Wolfram makes a compelling visual case that simple cellular automatalike rules might underlie such behavior in nature, and he gives a concise and readable review of some ideas in developmental morphology. But the results in this chapter do not advance beyond work done twenty or more years ago, and there are no new meaningful statements about any of these phenomena. Less compelling are Wolframs speculations on natural selection. He contends that because the generation of complexity is so common in simple cellular automata, it must be common in nature as well and, therefore, natural selection is not needed to create complexity in biology. In his view, biological systems are complex merely because evolution has sampled a huge number of simple programs and these often give rise to complex behavior. Wolfram offers no convincing evidence for this claim, nor does he discuss where these programs are implementedat the level of the genome? of cells? Although cellular automata provide plausible models of biological pigmentation and some aspects of morphology, there is as yet no compelling link between simple programs and complex biological systems such as the brain, the immune system, or cellular metabolism. On to capture the essence of what is going on even though traditional efforts have been quite unsuccessful. But as far as I can tell, his approach has yielded no new successful predictions, and none of his interesting speculations on physics propose any experiments to support this view.

Computational Universality A centerpiece of the book is Wolframs sketch of a proof done by Caltech graduate student (and former Wolfram research assistant) Matthew Cook that one of the elementary cellular automaton rules can support universal computation. In describing this result, Wolfram gives an excellent review of some central ideas Pigmentation patterns in theoretical computer scion mollusc shells. ence. The notion of univerWolfram interprets sal computation was first their striking resem- developed by Alan Turing blance to the patterns in the 1930s. Roughly, a produced by simple device is said to be uniImage not one-dimensional cellu- versal or can support lar automata as evi- universal computation if it available for dence that they are can run any program on online use. generated by processes any input. Nowadays, apwhose basic rules are proximations to a universal chosen at random from a set of the simplest devices are commonplace; they are known as propossibilities. grammable computers. The the contrary, it is increasingly clear that no- computer on your desk can (given enough tions of selection and adaptation are crucial memory) calculate any function, as long as for understanding such systems. the function is computable. (One of Turings greatest contributions was to demonFramework for Understanding Nature strate that noncomputable functions exist.) Since the beginning of the computer age, the In the early 1980s, Wolfram had found process of computation has been proposed as that of the 256 possible elementary celluan explanatory framework for many natural lar automaton rules (i.e., rules for onesystems. Artificial-intelligence practition- dimensional cellular automata with two ers have suggested that the brain is actually states and two neighbors per cell), a small a computer and that thinking is equivalent to subset, including the rule he numbered processing information. In the earliest use 110, exhibited particularly interesting beof cellular automata, von Neumann de- havior. Cook showed that, for any program scribed biological self-reproduction in com- and any input, one can specially design an putational terms (5, 6). More recently, all initial condition that encodes the program kinds of behavior (including the immune re- and input and then iterate rule 110 (of the sponse, the regulatory networks formed 256 possible rules for a one-dimensional among genes, and the collective behavior of cellular automaton with two states and two ants in a colony) have been cast as natural neighbors per cell) on the initial condition computation. Wolfram takes this notion of to, in effect, run the given program on the computation as a framework for nature to given input. Thus, rule 110 is a universal an extreme. He believes that nearly every- computer. This is an impressive result, and thing in the universe can be explained not Wolfram claims it is counterintuitive: just as computation but specifically in terms [I]t has almost always been assumed of simple programs such as cellular automathat in order to get something as sophistita. In a long, rather technical chapter, he discated as universality there must be no cusses how fundamental physics (thermodychoice but to set up rules that are themnamics, quantum mechanics, relativity, and selves special and sophisticated. One of the like) can be cast in terms of cellular authe dramatic discoveries of this book, tomata, a research program that was previhowever, is that this is not the case. ously pursued by Edward Fredkin, Tomasso Toffoli, Norman Margolus, Konrad Zuse, Wolfram views this accomplishment as and others [see (7)]. Wolfram claims that remarkably simple programs are often able extremely significant for science; he believes
VOL 298 SCIENCE www.sciencemag.org

66

4 OCTOBER 2002

CREDIT: STEPHEN WOLFRAM

S C I E N C E S C O M PA S S
65 64 63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 that a logical consequence is that universality is that it is easy to find universal computis ubiquitous throughout nature. Rule 110s ers even among simple cellular automata. universality will not, however, be very sur- The second claim is also plausible, though prising to computer scientists; Cooks proof supported by less evidence than the first. is only the latest in a series of demonstra- There are, in principle, processes that can tions that relatively simple devices (Turing compute things universal computers canmachines, neural networks, iterative maps) not, but these super-universal processes can be universal computers. Von Neumann require continuous-valued numbers (11). was the first to show that a cellular au- Whether continuous values actually exist tomaton can be universal. He constructed in nature and can be harnessed by natural a two-dimensional example with 29 states processes to surpass universal computaand four neighbors per cell, and others eventually reduced Image not the complexity to four states available for online use. per cell ( 8 ). In the 1970s, Conway sketched a proof that A universal computer. Rule 110 uses these eight functions to his Game of Life (with two specify a cells new color (lower row) for each possible combinastates and eight neighbors per tion of previous colors of the cell and its two neighbors (upper row). cell) is universal (2). Before Cooks work with rule 110, the simplest tion is unknown. Wolfram strongly beknown universal cellular automaton was lieves they cannot. The third claim does one-dimensional with seven states and two not make sense to me. I find it quite plauneighbors per cell (9). Rule 110 is now the sible that my brain is a universal computer simplest, and it is hard to see how a univer- (or would be, if I had infinite memory) sal cellular automaton could get any sim- and that the brain of the worm Caenorhabpler. Although it is interesting that such a ditis elegans is also (approximately) unisimple rule (not specifically constructed to versal, but I dont accept that the computaperform computation) turns out to be uni- tions we engage in are equivalent in soversal, the result is an incremental step over phistication. what had been done before. The significance of universality is also Summary tempered by practicality. Whereas rule 110 I think Wolfram is on the right track in (and other cellular automata) can be shown proposing that simple computer models to be universal in principle, in practice it is and experiments can lead to much almost impossible to design the initial progress. This approach may even come to condition necessary to perform a desired be seen as a new kind of science, though it computation. And even if such an initial will be the result of the contributions of a condition were known, the time needed to very large number of peoplebeginning perform the computation might be ex- with pioneers of the computer age such as tremely long compared with that on a tradi- von Neumann, Turing, and Norbert tional computer. Many people have claimed that the concepts of universal computation and uncomputability are relevant to science; in a notable example, Roger Penrose claimed these notions preclude the possibility of machine intelligence (10). But I know of no generally accepted scientific explanations or predictions in which these concepts play any role. Computational Equivalence The f inal chapter discusses Wolframs principle of computational equivalence, which is based on the idea that the best way to understand processes in nature is to view them as performing computations. It consists of three claims: (i) The ability to support universal computation is very common in nature. (ii) Universal computation is an upper limit on the sophistication of computations in nature. (iii) Computing processes in nature are almost always equivalent in sophistication. The first claim is plausible, though by no means established. Wolframs argument Wiener. (To be sure, Wolframs own contribution of the Mathematica software package has been of great value for such efforts.) I also agree that ideas from the field of computation will be increasingly useful for understanding natural phenomena, particularly in the study of living systems. Nonetheless, analytical approaches to illuminating complexity in nature thus far have been much more successful than cellular automata and related computational methods. A clear example is the successful use of the renormalization group to explain complex behavior in a wide range of dynamical systems (12). Given its length and content, A New Kind of Science is surprisingly readable. Wolframs use of pictures to illustrate difficult concepts works superbly well, and non-scientists will find it possible to understand much of what he covers. The principal obstacle readers face is the plethora of self-aggrandizement, some statements of which seem like they could not possibly be serious. For example, the author claims his principle of computational equivalence has vastly richer implications...than essentially any single collection of laws in science. Even more disturbing are the suggestions that Wolfram himself invented everything of interest here:
But to develop the new kind of science that I describe in this book I have had no choice but to take several large steps at once, and in doing so I have mostly ended up having to start from scratchwith new ideas and new methods that ultimately depend very little on what has gone before.

CREDIT: (TOP) STEPHEN WOLFRAM; (BOTTOM) IGOR BAKSHEE

Image not available for online use.

Artistic output. Igor Bakshee created this image using rule 110 and Mathematica.

In fact, most of what Wolfram describes is the work of many people (including himself), and most of it was done at least ten to twenty years ago. Nearly no credits to the contributions of others appear in the books main text. Some credits can be found in the long notes section at the books end, but many are not given at all. For example, the snowflake models Wolfram discusses are based on the work of Packard ( 13 ), but Packard is not mentioned in connection with them. This is only one example of such inexcusable omissions. Moreover, the book does not contain a single bibliographic citation

www.sciencemag.org

SCIENCE

VOL 298

4 OCTOBER 2002

67

S C I E N C E S C O M PA S S
65 64 63 62 61 60 59 58 57 56 55 54 53 52 51 50 49 48 47 46 45 44 43 42 41 40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 an astounding lapse that will put off serious scientific readers. Wolframs Web site (14) includes relevant books, but this list is no substitute. To benefit from the book, one must get past these issues without becoming too angry and take most of the claims with a large grain of salt. Wolframs discussions and speculations will interest many people in a wide variety of fields, but they do not constitute a new kind of science.
References and Notes
1. N. Metropolis, M. L. Stein, P. R. Stein, J. Comb. Theor. 15A, 25 (1973). 2. E. Berlekamp, J. H. Conway, R. Guy, Winning Ways for Your Mathematical Plays, vol. 2. (Academic Press, New York, 1982). 3. P. Prusinkiewicz, A. Lindenmayer, The Algorithmic Beauty of Plants (Springer-Verlag, New York, 1990). 4. E. N. Lorenz, The Essence of Chaos (Univ. of Washington Press, Seattle, WA, 1993). 5. J. von Neumann, A. W. Burks, Theory of SelfReproducing Automata (Univ. of Illinois Press, Urbana, IL, 1966). 6. A. W. Burks, Ed., Essays on Cellular Automata (Univ. of Illinois Press , Urbana, IL, 1970). 7. E. Fredkin, R. Landauer, T. Toffoli, Eds., Int. J. Theor. Phys. 21 (nos. 34, 67, 12), (1982). 8. E. R. Banks, dissertation, Massachusetts Institute of Technology (1971). 9. K. Lindgren, M. G. Nordahl, Complex Syst. 4 , 299 (1990). 10. R. Penrose, The Emperors New Mind (Oxford Univ. Press, 1989). 11. C. Moore, Theor. Comp. Sci. 162, 23 (1996). 12. M. J. Feigenbaum, Los Alamos Sci. 1 (no. 1), 4 (1980). 13. N. H. Packard, in Theory and Applications of Cellular Automata, S. Wolfram, Ed. (World Scientific, Singapore, 1986), pp. 305310. 14. www.wolframscience.com 15. I am grateful to R. Axelrod, D. Farmer, J. Holland, G. Huber, and C. Moore for discussions and insightful comments on an earlier version of the manuscript and to the Santa Fe Institute for support during the writing of this review.

B O O K S : E VO L U T I O N

Explanations for the Birds and the Bees


Paul Harvey

The author is in the Department of Zoology, University of Oxford, South Parks Road, Oxford OX1 3PS, UK. E-mail: paul.harvey@zoology.oxford.ac.uk

68

4 OCTOBER 2002

VOL 298

SCIENCE

www.sciencemag.org

CREDIT: JOE SUTLIFF

f I were an intellectually challenged adult male gorilla who stumbled across an adult male chimpanzee, I should in all likelihood be at a loss to explain my comparatively tiny testicles. Fortunately, my angst might be eased by consulting Dr. Tatiana, the agony aunt, who would point out that large testicles are characteristic of those primates and other mammalian species in which the female often mates with more than one male during a given estrus. Large testicles produce more sperm, thereby providing more tickets in the sperm competition lottery. Female go-

rillas mate only with the groups silverback evolutionary models. Through the device who, in the absence of sperm competition of having organisms describe their situafrom other males, needs to provide just tions (and predicaments) to her, she is able enough sperm to ensure that fertilization is to enter a dialog that uses individual case successful. The promiscuous female chim- studies to illustrate general principles. This panzee, on the other hand, has the sperm technique draws in the reader to a witty, from different males competing for access racy, informed, entertaining, and instructive to her eggs, so those males have evolved read. the capacity to produce inordinate quantiThere will be opposition to Judsons apties of the stuff. proach. Some will argue that anthropomorDr. Tatiana is the brainchild of Olivia phism on this level is unjustified and leads Judson, whose doctoral studies were su- inevitably to inaccuracies. Who cannot feel pervised by the late W. D. Hamilton. She for the plight of the green spoon worm wanted to describe to her audience our current understanding of the evolutionary biology of sex. The topic is manifold, wondrous, and Image not characterized by diversiavailable for ty: Why do some organisms have sexual reproonline use. duction whereas others do not? Why do different species have different numbers of sexes? What determines whether individuals are single-sexed or herma- ( Bonellia viridis ) that just inhaled her phrodite? Why do some species usually husband? But, then again, in what sense have imbalanced sex ratios while others was the male a husband before being indo not? Why is sex sometimes determined haled? Only after being inhaled does the genetically and sometimes environmen- male start to fertilize the females eggs. tally? What are the causes and conse- This example leads Dr. Tatiana to a descripquences of the different mating systems tion of environmental sex determination in seen in the natural world? Over the years the spoon worm: lone larvae mature into the variety has been described and the large females and larvae that subsequently problems of explaining it have been develop near a female become male. In a solved, to varying degrees. Many of the carefully crafted discourse that follows, she major contributions came from biologists explains why and when sex is environmenlike Darwin who became familiar with tally versus genetically determined. the natural history of many, many species It would be wrong to think about Sex and were then able to make comparisons Advice to All Creation as merely a collecto explain the differences. tion of anecdotes followed by descriptions Familiarity with natural history is of general principles. Instead, the book is a equivalent to becoming intimate with pri- developing text, meaning that it should be vate lives, except that the forread from the beginning bemer lacks the taboo of anthrocause answers to some quesDr. Tatianas pomorphism. Some of the best tions require familiarity with Sex Advice evolutionary biologists work by earlier chapters. For those who to All Creation attempting to identify themwant to check the facts for by Olivia Judson selves with the species they Chatto and Windus, themselves or to delve more study: What would I do if? is London, 2002. 317 pp. deeply into the problems that often useful shorthand for 16.99. ISBN 0-7011- Judson tackles, notes at the What would natural selection 6925-7. Metropolitan, end of the book cleverly referproduce under particular cir- New York, 2002. 319 pp. ence the original research pacumstances? Of course there $24, C$34.95. ISBN 0- pers used in its construction. can be dangers in this way of 8050-6331-5. The bottom line is that the thinking, which is why formal book actually works. Like models often reveal logical pitRichard Dawkinss Self ish falls. But, even then, the results of a logical Gene (Oxford University Press, Oxford, modeling process need to be described ver- 1976), it uses unabashed anthropomorbally. Judson has gone the whole hog by phism to create scenarios with which the employing anthropomorphism to its ex- open-minded reader can identify. Also like tremes, in the assurance that most of the Dawkins, Judson is a gifted writer, and her work she describes has been backed by book helps further understanding.